An annealing based approach to informed mode choice in agent-based transport simulations

Abstract

Modeling and simulating travelers mode choice is of great importance when evaluating novel transportation modes, policy measures or other changes to the overall transport system. The agent-based transportation framework MATSim has been widely used to study such scenarios and evaluate potential changes in the modal split. The default mode choice behavior in MATSim lets agents explore all their mode options randomly. A scoring-based co-evolutionary algorithm is applied to maximize the utility over a large number of iterations and arrive at consistent modes shares.In this paper, we present a new mode choice approach, that resembles a more rational decision-making process. First, the utilities of all available modes are estimated and used as costs to build a graph with all trips of the day. A k-shortest-path algorithm is applied to compute the most promising k combinations of modes for the day. Additionally, we formulate a pruning methodology that allows to remove unpromising modes. Finally, a selection model based on a multinomial logit is used to choose one of the plan candidates. The scale parameter of the selection model is annealed over the course of iterations to predict deterministically to the candidate with the highest estimate utility. Our simulation experiments show that convergence speed compared to random mode choice is significantly increased and only 25% of iterations are required to achieve overall better scores.